generating-semantic-game-mutants
Generating Semantic Game Mutants
Purpose
Measure what a test suite or an agent workflow actually catches, by planting defects whose ground truth you hold. Without planted defects, a suite's detection rate is unknown: green means "no defect was found", which is indistinguishable from "no defect was looked for".
The defects here are semantic and game-specific — a scoring guard removed, an identity reused
across a pool boundary, a field dropped from a save. Generic syntactic mutation (flip < to <=,
delete a statement) mostly produces crashes and dead code, which any suite catches, inflating the
score without measuring anything about gameplay.
What This Measures — Pick One Before Starting
These are different experiments with different setups. Running one and reporting another is the main way this instrument produces a wrong number.